[TTS] Refactor BOS and EOS handling to fix codec conversion

[rlangman]

This PR fixes a bug in MagpieTTS training in which self._codec_converter.convert_original_to_new() is being called on audio codec tokens after the Lhotse data loader has already pre-populated them with BOS and EOS tokens.

The addition of BOS and EOS tokens is now always done inside the model class, after convert_original_to_new() is called. These tokens are also removed before convert_new_to_original() is called, so that they are not mistakenly passed into the codec model.

There are also a few other minor changes, including:

• Fix data loading field mismatch when doing on-the-fly codec extraction (we now use target_audio instead of recording and context_audio instead of context_recording)
• Fix bug with inconsistent padding in spectral codec encoder
• Replace some manual padding done with torch.cat with simpler call to torch.nn.functional.pad

[blisc]

Closing, please make a new PR into main